Introduction to Bases in Banach Spaces

Matt Daws

June 5, 2005

Abstract We introduce the notion of Schauder bases in Banach spaces, aiming to be able to give a statement of, and make sense of, the Gowers Dichotomy Theorem.

Contents

1 Introduction 1

2 Schauder bases 2

3 Basic sequences 4

4 Unconditional bases 6

5 The Gowers Dichotomy Theorem 9

1 Introduction

When working with finite dimensional vector spaces, it is often convenient to take a basis, and then work with the co-ordinate system which this basis gives us. In an infinite dimensional vector space, the axiom of choice shows that we can still find a basis: however, because it is now infinite, its use is often less. This is especially true in a Banach space, because such an algebraic basis takes no account of the extra structure induced by the norm. This leads to the notion of a Schauder basis. Let E be a Banach space and let (en) be a sequence of vectors in E. Then (en) is a Schauder basis (or, from now on, simply a basis) if every x ∈ E admits an expansion of the form

∞ X x = xnen, n=1

for some unique sequence of scalars (xn). Note. As someone interested in algebraic questions, I always work over the complex numbers. However, it is common (for simplicity) to work with real scalars in Banach space theory. Often, a similar result for complex numbers follows easily, sometimes we have to work somewhat harder, and occasionally, the result is true only for real scalars. We shall try to give proofs which work with either choice of scalar, but occasionally the reader should assume the result is only certainly known for real scalars.

1 p p Example 1.1. Let E = l for 1 ≤ p < ∞, or E = c0. Note that some authors write lp, ` , or `p instead. For n ≥ 1, let en ∈ E be the sequence which is 0 except with a 1 in the nth position. Then (en) is a Schauder basis for E, called the “standard unit-vector basis” of E.

Example 1.2. Let H be a separable Hilbert space with an orthonormal basis (en). Then (en) is a basis (in fact, an unconditional basis: see later) for H.

2 Schauder bases

We ﬁrst give some general results on Schauder bases. Let E be a Banach space with a basis (en). Clearly, we are free to multiply each en by a non-zero scalar without changing the basis property. Hence we may assume that (and will henceforth do so) (en) is a normalised basis, that is, kenk = 1 for each n. Then, for x ∈ E, we can uniquely write P x = xnen, and hence deﬁne

N X kxk0 = sup xnen . N≥1 n=1 PN This is well-deﬁned, as n=1 xnen → x as N → ∞, and every convergent sequence in a Banach space is clearly bounded. We also easily see that kxk ≤ kxk0 for each x ∈ E, and that kxk0 = 0 if and only if x = 0.

Theorem 2.1. Let E be a Banach space with a basis (en), and let k · k0 be as deﬁned above. Then k · k and k · k0 are equivalent norms on E.

Proof. Let F be the normed space which is E together with the norm k · k0 (it is easily checked that F is indeed a normed-space). Let ι : F → E be the formal inclusion map. Then ι is norm-decreasing, and is a bijection. Suppose that F is Banach space, that is, F is complete. Then the Open Mapping Theorem implies that ι has a continuous inverse, which implies that k · k0 is equivalent to k · k. P∞ So, we wish to show that F is complete. Firstly, given x = n=1 xnen, and m ≥ 1, we notice that m m−1 −1 −1 X X |xm| = kxmemkkemk = kemk xnen − xnen n=1 n=1 m m−1 −1 X X −1 ≤ kemk xnen + xnen ≤ 2kemk kxk0, n=1 n=1 by the deﬁnition of k · k0. Now let (xn) be a Cauchy-sequence in F , and let ∞ X xn = xn,mem (n ≥ 1). m=1

Then, for > 0, there exists N so that for r, s ≥ N,

N X kxr − xsk0 = sup (xr,m − xs,m)em < . N≥1 m=1 By the calculation above, we hence see that for each m ≥ 1, the sequence of scalars ∞ P (xn,m)n=1 is a Cauchy-sequence, with limit ym say. We now show that y = ymem

2 converges in F . As k · k0 ≥ k · k, this implies that the sum converges in E, and hence that y exists, as E is complete. Towards this end, let n ≥ N, so that for each N ≥ 1, we have that N N N X X X (ym − xn,m)em = lim (xr,m − xn,m)em = lim (xr,m − xn,m)em ≤ , r→∞ r→∞ m=1 m=1 m=1 by the deﬁnition of N. Notice that this shows that y is indeed the limit of (xn), supposing P∞ of course that y exists. Then m=1 xn,mem converges in E, so there exists M such that Ps if M ≤ r < s, then m=r xn,mem < . Thus we have that

s min(t,s) X X ymem = sup ymem 0 t≥r m=r m=r t t X X ≤ sup (ym − xn,m)em + xn,mem r≤t≤s m=r m=r t r−1 X X < sup (ym − xn,m)em + (ym − xn,m)em + ≤ 3, r≤t≤s m=1 m=1 P so that m ymem does converge in F .

We can now give a “ﬁnite” characterisation of a basis.

Theorem 2.2. Let E be Banach space, and let (en) be a sequence in E. Then (en) is a basis for E if and only if:

1. each en is non-zero;

2. the linear space of (en) is dense in E;

3. there exists a constant K such that for every sequence of scalars (xn), and each N < M, we have that N M X X xnen ≤ K xnen . n=1 n=1

Proof. Suppose that (en) is a basis, and form the norm k · k0, as above, so that by the above theorem, there exists K such that

kxk ≤ kxk0 ≤ Kkxk (x ∈ E). P∞ Clearly we have condition (1) by the uniqueness of the expansion of 0 = n=1 0en; condition (2) is trivial. Then, for (3), let (yn) be the sequence of scalars deﬁned by yn = xn for n ≤ M, and yn = 0 otherwise. Then we see that N N ∞ ∞ M X X X X X xnen = ynen ≤ ynen ≤ K ynen = K xnen , 0 n=1 n=1 n=1 n=1 n=1 as required. Conversely, let F be the linear span of (en), a dense subspace of E. For each n ≥ 1, deﬁne a linear map Pn : F → F by N n N X X X Pn xkek = xkek N ≥ n, xkek ∈ F . k=1 k=1 k=1

3 By condition (3), we see that kPnk ≤ K. It is then clear that Pn is a bounded projection on F , and that Pn thus extends to a bounded projection on E. For each n ≥ 1, deﬁne a ∗ linear functional en on F by the formula

∗ en(x)en = Pn(x) − Pn−1(x)(x ∈ F ),

∗ ∗ where P0 = 0. Then en is well-deﬁned upon F (by condition (1)) and kenk ≤ 2K, so that ∗ en extends by continuity to a bounded linear functional on E. It is clear that

n X ∗ Pn(x) = ek(x)ek (x ∈ F ), k=1 so by continuity, this formula holds for x ∈ E as well. Now let x ∈ E. We claim that

∞ X ∗ x = en(x)en. n=1

Let (xn) be a sequence in F which converges to x. Then, for > 0, let M ≥ 1 be such that kx − xM k < , and let N ≥ 1 be suﬃciently large so that PN (xM ) = xM , which we can do, as xM ∈ F . Then we see that

N X ∗ en(x)en − x = kPN (x) − xk n=1

≤ kPN (x) − PN (xM )k + kPN (xM ) − xM k + kXM − xk

≤ Kkx − xM k + kxM − xk < (K + 1).

Finally, we note that if ∞ ∞ X X ∗ x = xnen = en(x)en, n=1 n=1 ∗ ∗ then applying eN , we conclude that xN = eN (x) for each N, so that such an expansion is unique.

The smallest constant which can arise in (3) is the basis constant of the basis (en). Theorem 2.1 shows that we can always renorm E to give a basis on 1. Such bases are called monotonic. It is immediate that a Banach space with a basis is separable, so that, for example, l∞ does not have a basis. Most common separable Banach spaces do have bases, although they are often non-obvious to find. For a while, it was thought that all separable Banach spaces would have a basis, but Enflo produced a counter-example in [Enflo, 1973]. There is, however, a weaker notion where we do have a positive result.

3 Basic sequences

Let E be a Banach space, and let (xn) a sequence in E such that (xn) is a basis for its closed linear span. Then (xn) is a basic sequence in E. It is easy to show that every Banach space has a basic sequence; that is, every Banach space E contains a closed, inﬁnite-dimensional subspace F with a basis.

4 Lemma 3.1. Let E be an inﬁnite-dimensional Banach space, let F be a ﬁnite-dimensional subspace of E, and let > 0. Then there exists x ∈ E such that kxk = 1 and

kyk ≤ (1 + )ky + axk,

for all y ∈ F and all scalars a.

Proof. We may suppose that < 1. As the unit ball of F is compact, there is a ﬁnite set {y1, . . . , yn} in F such that

kykk = 1 (1 ≤ k ≤ n), min ky − ykk < /2 (y ∈ F, kyk = 1). 1≤k≤n

∗ ∗ 0 ∗ Pick y1, . . . , yn norm-one vectors in E , the dual of E, such that yk(yk) = 1 for each k. ∗ Then there exists x ∈ E with kxk = 1 and yk(x) = 0 for each k. For any norm-one y ∈ F , pick yk such that kyk − yk < /2. For a scalar a, we have that

ky + axk ≥ kyk + axk − ky − ykk > kyk + axk − /2 ∗ −1 ≥ |yk(yk + ax)| − /2 = 1 − /2 ≥ (1 + ) , as required.

Theorem 3.2. Every Banach space E contains a basic sequence.

Proof. We use induction to pick a sequence of norm-one vectors (xn) such that condition (3) of Theorem 2.2 always holds, with K = 2 say (the proof works for any K > 1). Suppose we have chosen x1, . . . , xn and > 0 such that

m n X X akxk ≤ (2 − ) akxk , k=1 k=1

n for any m ≤ n and any scalars (ak)k=1. Note that we can clearly do this for n = 1. We now try to ﬁnd xn+1. We need to ensure that kxn+1k = 1 and that for some 0 > 0, we have that m n+1 X X akxk ≤ (2 − 0) akxk , k=1 k=1

n+1 for any m ≤ n and any scalars (ak)k=1. Let Fn be the linear span of x1, . . . , xn, a ﬁnite-dimensional subspace of E. Use the above lemma to ﬁnd a norm-one vector xn+1 such that kyk ≤ (1 + δ)ky + an+1xn+1k for each y ∈ Fn and each scalar an+1, where δ > 0 is chosen so that (2 − )(1 + δ) = 2 − /2, n+1 Pn that is, δ = /2(2 − ). Then, for a sequence of scalars (ak)k=1, let y = k=1 akxk ∈ Fn, so that for m ≤ n, we see that

m n X X akxk ≤ (2 − ) akxk = (2 − )kyk ≤ (2 − )(1 + δ)ky + an+1xn+1k k=1 k=1 n+1 X = (2 − /2) akxk , k=1 as required.

5 4 Unconditional bases

A basis for a Banach space is not, in many cases, of a huge amount of use, as the convergence properties are rather weak. A more useful notion is that of an uncondtional basis. A series (xn) in a Banach space E is said to sum unconditionally if, for each permutation σ on N, the sum ∞ X xσ(n) n=1 converges, and converges to the same limit (although, see below, this is automatic), independently of σ. Recall that in a finite-dimensional Banach space (or just in R or C) P∞ this notion is equivalent to n=1 kxnk < ∞. This is not true in the infinite-dimensional case (indeed, it characterises finite-dimensional Banach spaces). Then a basis (en) for a Banach space E is an unconditional basis if, for each x ∈ E, there exists a unique expansion of the form

∞ X x = xnen, n=1 where the sum converges unconditionally. Firstly, we need to explore some properties of unconditional convergence in Banach spaces.

Proposition 4.1. Let E be a Banach space, and let (xn) be a series in E. Then the following are equivalent: P∞ 1. the sum n=1 xn is unconditional; P∞ 2. for each permutation σ of N, the sum n=1 xσ(n) converges in E;

3. for each > 0, there exists a finite subset A ⊆ N such that if B ⊆ N is finite with P A ∩ B = ∅, then k n∈B xnk < . Proof. Clearly (1) implies (2). To show that (2) implies (3), suppose towards a contradiction that (3) does not hold, so that we can find > 0 such that for each finite A ⊆ N, there exists a finite B ⊆ with A ∩ B = ∅ and k P x k ≥ . So let B = B , A N A n∈BA n 1 {1} then let A2 = {1, 2,..., max B1}, and let B2 = BA2 . Continue in this manner to find a sequence of finite sets (Bn) such that max Bn < min Bn+1 for each n ≥ 1. We can easily construct a permutation σ of N such that each Bn is the image under σ of some interval, say Bn = {σ(rn), σ(rn + 1), . . . , σ(sn)}, so that sn X X ≤ xk = xσ(k) (n ≥ 1). k∈Bn k=rn P However, this implies that k xσ(k) does not converge, giving the required contradiction. Now suppose that (3) holds, and that σ is a permutation of N. For > 0, choose −1 A ⊆ N and let N > 0 be such that if n ≥ N, then σ(n) 6∈ A, σ (n) 6∈ A, and n 6∈ A. Then, for N ≤ r < s, let B = {σ(n): r ≤ n ≤ s} and let C = {n : r ≤ n ≤ s}, so that B and C are finite, and B ∩ A = C ∩ A = ∅. Thus we have that s s X X X X xσ(n) = xn < , xn = xn < , n=r n∈B n=r n∈C

6 P P which shows that x = n xn and y = n xσ(n) exist. Furthermore, let K = {1 ≤ n ≤ N : σ(n) > N} and L = {σ(n) : 1 ≤ n ≤ N, σ(n) > N}, so that L ∩ A = ∅. Also, if −1 n ∈ K, then m = σ(n) > N so that σ (m) = n 6∈ A, that is, K ∩ A = ∅. Then we see that N N X X X X xn − xσ(n) = xn − xn < 2, n=1 n=1 n∈K n∈L which easily shows that kx − yk ≤ 4. Thus we see that (1) holds. P Notice that condition (3) shows that if xn converges unconditionally, then given P n any A ⊆ N, the sum n∈A xn converges unconditionally. Indeed, we can say more.

Proposition 4.2. Let E be a Banach space, and let (xn) be a series in E. If (xn) sums P unconditionally, then for each sequence (zn) in {+1, −1}, the sum n znxn converges uniformly, independently of the choice of (zn). Furthermore, the following are equivalent: P∞ 1. the sum n=1 xn is unconditional; P∞ 2. for any bounded sequence of scalars (zn), the sum n=1 znxn converges in E. P Proof. We use condition (3) from the above proposition to characterise when xn P n converges unconditionally. So suppose that x = xn converges unconditionally. By n P the above comment, for each A ⊆ N, we may define PA(x) = xn. Then suppose n∈A P that (zn) is a sequence in {+1, −1}, so that if A = {n : zn = 1}, then clearly n znxn = PA(x) − PN\A(x) converges. Suppose that there exists > 0 such that for each r ≥ 1, there exists a sequence (an) Ps in {+1, −1} and s > r such that k n=r anxnk ≥ . Then we find can sequences (An) and (Bn) of finite subsets of N such that An ∩ Bn = An ∩ Am = Bn ∩ Bm = ∅ for all n 6= m, and

X X xk − xk ≥ . k∈An k∈Bn

Let σ and τ be permutations of N such that each An occurs as the image under σ of P an interval, and each Bn occurs as the image under τ of an interval. As n xσ(n) and P x converge, we again conclude that for n suﬃciently large, k P x k < /3 n τ(n) k∈An k and k P x k < /3, which is a contradiction. This shows that P z x converges k∈Bn k n n n uniformly, independently of the choice of the (zn). Suppose now we are working over the real numbers, and let (zn) be a bounded sequence Ps in R. For 1 ≤ r < s, pick µ in the dual space to E with kµk = 1 and n=r znµ(xn) = PN k n=1 znxnk. For each n, let an = 1 if µ(xn) ≥ 0 or an = −1 otherwise, so that s s s s X X X X znxn ≤ |zn||µ(xn)| = |zn|anµ(xn) ≤ k(zn)k∞ anxn . n=r n=r n=r n=r Ps By the above, we know that we can make k anxnk arbitrarily small, independently Pn=r of the choice of (an). Thus we conclude that n znxn converges. We have already essentially shown that (2) implies (1), just by taking sequences (zn) in {+1, −1}.

We are now in a position to prove a version of Theorem 2.2 for unconditional bases: this is left as an exercise to the reader. Instead, we prove the hard parts of such a theorem.

7 Theorem 4.3. Let E be a Banach space, and let (en) be an unconditional basis for E. P For x = n xnen, deﬁne ∞ X kxk0 = sup znxnen . k(zn)k∞≤1 n=1

Then k · k0 is an equivalent norm on E.

Proof. As (en) is a basis for E, we may assume that (en) is actually a monotone, normalised basis. Firstly, note that the above proposition shows that k · k0 is indeed well- deﬁned. This follows as for > 0, there exists N, independent of (zn), such that if Ps P N ≤ r < s, then k n=r znxnenk ≤ k(zn)k∞. Thus we have that k n znxnenk ≤ PN k(zn)k∞ + |xn|kenk . n=1 P We proceed as in the proof of the preceeding proposition. For x = n xnen, deﬁne ∞ X kxk1 = sup znxnen , (zn)∈{+1,−1}N n=1

again, this is well-deﬁned. If we can show that k · k1 is an equivalent norm on E, then the argument used above shows that k · k0 is also an equivalent norm on E. Let Λ be the collection of sequences in {+1, −1}, which is a group under pointwise multiplication. For each z = (zn) ∈ Λ, let Tz : E → E be the linear map

∞ ∞ X X Tz xnen = znxnen. n=1 n=1 P Suppose that (xn) is a sequence in E tending to 0, that xn = xn,mem for each n, and P m that Tz(xn) → m ymem as n → ∞. That is, ∞ X lim (zmxn,m − ym)em = 0. n→∞ m=1

∗ As (en) is a monotone basis, we may apply the co-ordinate functional em to conclude that

lim |zmxn,m − ym| = 0 (m ≥ 1). n→∞

∗ However, as xn,m = em(xn) → 0 as n → ∞, we see that ym = 0 for each m, so that y = 0. The Closed Graph Theorem then implies that Tz is bounded. The above proof shows that, for a ﬁxed x ∈ E, the family {Tz(x): z ∈ Λ} is bounded. It follows from the Principle of Uniform Boundedness that the family of operators {Tz : z ∈ Λ} is bounded. However, this is simply the statement that k · k1 is equivalent to k · k, as required.

∞ Example 4.4. In c, the closed subspace of l consisting of convergent sequences, let fn = (0,..., 0, 1, 1,...) ∈ c, where the 1 appears in the nth position. Then, for x = (xn) ∈ c, let y1 = x1, and yn = xn − xn−1 for n ≥ 2, so that

N X ynfn = y1, y1 + y2, . . . , y1 + y2 + ··· + yn,... n=1 = x1, x2, . . . , xn, xn,... ,

which tends to x as N → ∞. That is, (fn) is a basis for c, called the summing basis.

8 However, we notice that

N N X X n fn = N, (−1) fn = 1, n=1 n=1 so that the above theorem implies that (fn) is a conditional basis.

Let E be a Banach space with an unconditional basis (en). Then we may renorm E so that there is a norm-decreasing algebra homomorphism l∞ → B(E), from the Banach algbera l∞ under pointwise multiplication to the algebra of operators on E. P In particular, for A ⊆ N, deﬁne an operator PA : E → E by PA( xnen) = P n n∈A xnen. Then {PA : A ⊆ N} is a bounded family of projections on E; by renorming, we may suppose that each PA is actually norm-decreasing.

5 The Gowers Dichotomy Theorem

Let E be a Banach space, and let P be a bounded projection on E. Then let F be the kernal of P , a closed subspace of E, and let G be the image of P , which is equal to the kernal of I − P , and is hence a closed subspace. It is trivial to see that E = F ⊕ G. Conversely, if E = F ⊕ G for some closed subspaces F and G, then for each x ∈ E, we can write x = f + g for unique f ∈ F and g ∈ G. It is easy to see that the map x → g defines a bounded projection on E with kernel F and image G. If F is a finite-dimensional subspace of E, then by taking a finite basis of F , we can easily define a bounded projection E → F . Hence E always admits some projections: however, need there exist a projection with an infinite-dimensional image? An infinite-dimensional Banach space E is indecomposable if whenever we can write E as the direct sum of two closed subspaces, E = F ⊕ G, then one of F or G is finite- dimensional. That is, the only bounded projections on E are the trivial ones. A Banach space E is hereditarily indecomposable if, furthermore, every infinite-dimensional closed subspace F of E is indecomposable. That such Banach spaces exist is somewhat surprising, but in [GM, 1993], Gowers and Maurey constructed an example of such a space. Notice that by the comments at the end of the last section, a hereditarily indecomposable Banach space E cannot contain an unconditional basic sequence, or we would have an infinite dimensional closed subspace F which would admit many non-trivial bounded projections. Indeed, this was the problem which Gowers and Maurey were trying to find a counter-example to. Hence we see that being hereditarily indecomposable and having an unconditional basic sequence are somehow the opposite of each other. Remarkably, this statement can be made precise, in the Gowers Dichotomy Theorem. Theorem 5.1. Let E be a Banach space. Then there exists a subspace F of E such that either F is hereditarily indecomposable, or F has an unconditional basis. Proof. See [Gowers, 1996] for a discussion, and [Gowers, 2002] for the full proof.

References

[Enﬂo, 1973] Enflo, P., ‘A counterexample to the approximation problem in Banach spaces’, Acta Math. 130 (1973) 309–317. [Gowers, 1996] Gowers, W.T., ‘A new dichotomy for Banach spaces’, Geom. Funct. Anal. 6 (1996) 1083–1093.

9 [Gowers, 2002] Gowers, W. T., ‘An inﬁnite Ramsey theorem and some Banach-space dichotomies’, Ann. of Math. 156 (2002) 797–833. [GM, 1993] Gowers, W. T., Maurey, B., ‘The unconditional basic sequence problem’, J. Amer. Math. Soc. 6 (1993) 851–874.